Repository navigation
Prune is too aggressive for long-context models and should scale with model context #16285
Description
Problem
SessionCompaction.prune() currently uses fixed thresholds:
PRUNE_PROTECT = 40_000PRUNE_MINIMUM = 20_000
and runs after each assistant turn.
That means old tool outputs start being replaced with [Old tool result content cleared] after only ~40K protected tool-output tokens, regardless of whether the active model has a 128K, 200K, or 1.05M context window.
Why this feels wrong on 1M-context models
Full compaction already looks at model limits, but prune does not. So on long-context models, prune can start clearing old tool results very early even when the session is still nowhere near the actual compaction / overflow threshold.
In practice this feels like premature memory loss on large-context models: the model still has plenty of room left, but important older tool output is already gone.
Current workaround
There is a boolean escape hatch via compaction.prune: false, but that is all-or-nothing. There is no way to tune the thresholds.
Request
Could you consider one or more of the following?
- scale prune thresholds from
model.limit.context - add config for thresholds such as
prune_protect/prune_minimum - make prune less aggressive automatically for very large-context models
- expose clearer logs / UI around what was pruned and why
I think this would make long-context models behave much more intuitively.
Activity
github-actions commentedon Mar 6, 2026
This issue doesn't fully meet our contributing guidelines.
What needs to be fixed:
[FEATURE]:and the feature request issue template was not used. Please resubmit using the feature request template, which requires a verification checkbox and description, with the title prefixed by[FEATURE]:.Please edit this issue to address the above within 2 hours, or it will be automatically closed.
This issue might be a duplicate of existing issues. Please check:
thresholdandmaxContextconfig optionsIf you believe this was flagged incorrectly, please let a maintainer know.
nyl199310 commentedon Mar 6, 2026
Closing this in favor of #11930, since that issue is already tracking configurable compaction behavior.
I added a follow-up comment there with the prune-specific details from this issue, because my main concern is that
prunecurrently uses fixed thresholds and feels too aggressive on long-context models.